We ran the Can Bots scan against every Winter 2026 Y Combinator company website to see how well AI agents and LLMs can actually read them. Here's what the machines see.
Most company sites still flunk the agent test. Winter 2026 does better than the YC historical average — but 79 of 187 sites still score an F.
Average score across every scanned Winter 2026 company, per dimension, as a share of that dimension's max.
Semantics and Security are the batch's strengths. Agent Protocols (llms.txt, MCP, machine-readable action hints) is the weakest at 5.5/15 — the clearest place to win points.
The highest-scoring Winter 2026 companies — sites already built for an agent-first web.
| Company | Grade | Score | Disc | AI | Sem | Agent | Sec |
|---|
The Can Bots scan checks whether a crawler can fetch your pages, read your structure, understand your content, and find machine-readable signals. A passing score means a bot can see you.
It can't tell you whether an agent can actually use your product — sign up, complete a flow, or get a real answer. That takes a human-style task run, which is a different test entirely.
The short version: passing the scan gets you into the room. It doesn't mean the agent can do anything useful once it's there.
Cleaned up your scan score? Take it further. YC-backed founders get 15% off a SaaS Dummies report — real AI testers that actually use your product and show you exactly where people get stuck.
Apply YCF25 at checkout.
The Can Bots scan is the static check. Run your own free grade across all five categories, then see how real testers move through your product.